search

YanjieZe / Improved-3D-Diffusion-Policy

[arXiv 2024] Generalizable Humanoid Manipulation with Improved 3D Diffusion Policies. Part 1: Train & Deploy of iDP3
MIT License
139 stars 12 forks source link

About the improved visual encoder #8

Closed JLHins closed 2 weeks ago

JLHins commented 2 weeks ago

Hi, thanks for sharing the great work.

The results in this paper demonstrate a performance boost with a stronger visual encoder, but in the previous DP3 we get the opposite. How should we understand the difference here?

YanjieZe commented 2 weeks ago

Hi, thank you for your interest. If you mean the ablation exp in DP3, I think the visual encoder of iDP3 is different from the ones ablated in DP3. We are not showing a strong encoder is not good for DP3, but showing most other encoders are not suitable.